ChatProductHow it worksLimitsPlatformsModelsPricingCompanyCareersLegalTeam

By Raman, Jayson, Ashly, and Emily

Beta 3.8 · public release Q4 2027

Formerly computer.vodka. Now Subterminal Agents.

September 2026
Models you already run, plus Apple

Point the harness at Grok, Gemini, OpenAI, Claude, Qwen, DeepSeek, Llama, Mistral, Ollama, and any OpenAI-compatible local server. Same install on macOS, iOS, and watchOS.

See every month since January 2026
Grok Gemini OpenAI Claude Ollama Llama Mistral QwenQwen DeepSeek

An agent harness that runs on the machine you already use.

The control loop around the model you already run. Plan the turn, dispatch one tool, read stdout or the file diff, then stop or wait on an approval gate. Tools, memory, and the session log live in one workspace folder.

Request beta 3.8 What it can and can't do
Windows Apple Linux Android

Same harness on Windows, Linux, Android, and the Apple ecosystem: macOS, iOS, and watchOS.

session probe

One harness. Point it at the model you already run.

Grok Gemini OpenAI Claude Ollama Llama Mistral QwenQwen DeepSeek Any /v1 endpoint

The harness, in a chat

You type the task. The harness runs the loop.

Plan, one tool, observe, stop. Writes, deletes, installs, and outbound sends wait for a yes. After setup, prompts and files do not go to our servers.

Approval before damage

You see the plan and the diff before anything hits disk. Paths resolve against one folder root.

Your model, your folder

Ollama, llama.cpp, LM Studio, vLLM, or any OpenAI-compatible endpoint on localhost or the LAN.

Memory is markdown

Notes are .md files you can open or delete. Optional local embeddings stay on disk. Plain transcript.

Subterminal Agents
Preview · cannot see your disk
beta 3.8

Apps as switches

How it works

Plan, tool, observe, approve.

The same loop on every screen we ship, including a watch glance so you can confirm a write from the wrist.

01. Plan

Hand off a task

Point the harness at a folder and a model. It plans the turn and picks one tool. Nothing writes yet.

02. Observe

Read the output

Stdout, a file diff, a failed test. The next turn only starts after the harness has seen the result.

03. Gate

Yes before damage

Deletes, overwrites, installs, and outbound sends wait. Keep your own backups. No built-in git.

Platforms

Same loop on the machines you already own.

Desktop

Windows, Linux, macOS

Folder manager on the desk. Local or LAN model. Works offline after setup.

Phone

iOS and Android

Same chat shell. SMS and WhatsApp after a password check on the first message.

Wrist

watchOS

Glance a task. Approve a write from the wrist. Same gate as the laptop.

Limits

What it can and can't do.

The harness is not a branded brain. If the weights are thin, the answers are thin.

It can

✓
Keep working offline after setup

Once the binary and model are on the machine, prompts and files do not go to our servers.

✓
Run across machines you already own

Windows, Linux, Android, macOS, iOS, watchOS. Same loop, plus optional SMS and WhatsApp.

✓
Talk to connected apps

Mail, calendar, docs, chat, drive. Off means the tool is never offered.

✓
Pick up a task from a text

SMS or WhatsApp after a password check. Hosted gateway is a subscription extra.

✓
Research, summarise, extract

Open a page you name. Grind a folder of PDFs, CSVs, or code inside the workspace root.

✓
Code, debug, and test in the folder

Read a repo, patch a file, run the test you approve, paste the failure into the next turn.

✓
Draft mail, briefs, and reports

Write the thing, show it, wait for send. Connected inboxes are a connector.

✓
Remember across turns on disk

Plain markdown notes plus optional local embeddings. Delete a memory by deleting the file.

It can't, at least not yet

–
Generate or edit video

It can organise clips and write ffmpeg plans you approve.

–
Match frontier cloud models on the free path

A 7B or 14B on a laptop is not GPT-class. Hosted extras exist for hard turns.

–
Run well on a weak machine

Useful local models need RAM and, ideally, a GPU.

–
Browse the open web by default

Network fetch is off until you turn it on, and only for pages you name.

–
Always read the task correctly

It can mis-scope a folder. Review the diff before you approve.

–
Replace git or backups

There is no built-in version control.

–
Act from this website

The chat on this page is a preview. The loop starts after you install beta 3.8.

–
Be a single branded brain

Subterminal Agents is the harness, not the model.

Pricing

The harness stays free. Hosted extras are usage-only.

Install it, point it at your model, work in a folder. No card. From Q4 2027, hosted extras are optional and metered.

Always free

Stays local

No subscription · ships in beta 3.8 · quality = your model

✓
The full harness loop

Plan, tool, observe, stop. Read, write, search, shell, local fetch inside the folder you choose.

✓
Your own model

Ollama, llama.cpp, LM Studio, or any OpenAI-compatible endpoint.

✓
Memory and logs on disk

Markdown notes and a plain transcript. No account.

✓
Approval before damage

Deletes, overwrites, and installs wait for a yes. That gate does not move to the cloud.

✓
Works offline after setup

No seat, no monthly fee, no phone-home just to keep using one computer.

Early access · monthly

Hosted extras

$69.67 / month

Early paid build. $69.67 each month with a usage cap. Not unlimited. Per-token metering starts in Q4 2027.

✓
Cloud model when local is not enough

Included allowance each month. Files stay unless you attach them. True per-use metering arrives with the Q4 2027 public beta.

✓
Same session on more than one device

Sync chat and memory notes across a laptop and a phone.

✓
Always-on text and WhatsApp

Hosted gateway when the lid is closed.

✓
Live app connectors

Gmail, Calendar, Slack, Drive need an OAuth broker we host.

✓
Jobs that outlive sleep

A hosted runner keeps a long task going after the lid closes.

$69.67 is billed every month. Usage is capped. It is not unlimited, and it is not billed per token. Per-token metering starts with the Q4 2027 public beta. This early build is unstable. Crashes, lost files, broken connectors, and unfinished features are expected. By requesting access you accept that Subterminal Agents takes no liability for what happens on your machine, files, accounts, or time. The free local install stays the supported path.

About us

Small team. Local runtime.

Four people. One product. If a model farm wraps an API and calls it an agent, that is not this.

R

Raman

Founder · Runtime and inferenceWellington, New Zealand

Owns the loop that actually runs: plan, one tool, read stdout, stop. Wires Ollama, llama.cpp, LM Studio, vLLM, and any OpenAI-style /v1 endpoint. Ships the install on Windows, Linux, Android, and Apple Silicon.

Rust / C++OpenMPLocal inferenceStreaming HTTP
[email protected]
J

Jayson

Co-founder · Product and evaluationAustralia

Holds the ship date. Writes folder evals that finish a real task, blocks a release if a gate failed, and keeps hosted extras metered. If a demo only works on one laptop, it does not go out.

Eval harnessesRelease barPricingQA traces
[email protected]
A

Ashly

Co-founder · Design and interfaceUnited Kingdom

Designs the shell you tap: chat, capability switches, approval states, desktop, phone, and the watch glance. Type, motion, and the >_ mark. If a control is loud, it does not ship.

Product UIMotionFigmaSwiftUI / CSS
[email protected]
E

Emily

Co-founder · Memory and tool safetyUnited States

Memory is markdown in the workspace. Every path resolves against one folder root. Deletes, overwrites, installs, and outbound sends wait for a yes. First-message password check on SMS and WhatsApp.

Path sandboxApprovalsKeychainTranscripts
[email protected]

Company

A small studio. We ship a runtime, not a model farm.

What we sell

The install is free. Quality follows the model you attach. Hosted extras from Q4 2027 are optional and metered.

How we work

Remote-first across New Zealand, Australia, the United Kingdom, and the United States.

Contact

Beta, press, partnerships, and security all go to [email protected]. No form first.

Beta 3.8 open Public beta Q4 2027 NZ · AU · UK · US [email protected]

Updates

One note a month since we started building.

Public log. Newest first. Beta 3.8 is open to request.

  1. September 2026 · current

    Models you already run, plus Apple

    Grok, Gemini, OpenAI, Claude, Qwen, DeepSeek, Llama, Mistral, Ollama, any OpenAI-compatible local server. Same install on macOS, iOS, watchOS.

    GrokwatchOSCareers
  2. August 2026

    Pricing that does not tax the disk

    Free local stays the full loop. Hosted extras from public beta are usage-only. No seat fee.

  3. July 2026

    Apps as switches, not a plugin zoo

    Mail, calendar, docs, and chat sit behind the Apps toggle. Off means the tool is never offered.

  4. June 2026

    Text the harness

    SMS and WhatsApp drop a task into the same loop. First message is password-checked.

  5. May 2026

    Attach whatever you already run

    Ollama, llama.cpp, LM Studio, vLLM, or any OpenAI-compatible endpoint.

  6. April 2026

    Windows, Linux, and Android installs

    Same loop on every desktop and phone we could test. No account after setup.

  7. March 2026

    Approval before damage

    Deletes, overwrites, installs, and outbound sends wait for a yes.

  8. February 2026

    Memory is markdown in the folder

    No hidden vector store. Notes are .md files. Every tool call lands in a plain transcript.

  9. January 2026

    The loop, on one machine

    Plan, pick one tool, read the result, stop. Four people, one product. Beta 3.8 first.

Careers

Open roles

Remote across NZ, AU, UK, and US. Pick a role, then skip or apply. Write to [email protected].

See the roles again

You have been through every open seat.

Email careers@

Apply

Beta 3.8

Request the private build.

Public beta stays Q4 2027 until the install path is stable. We’ll draft the request to [email protected].